HyperNST: Hyper-Networks for Neural Style Transfer

We present HyperNST; a neural style transfer (NST) technique for the artistic stylization of images, based on Hyper-networks and StyleGAN2 architecture. Our contribution is novel method inducing parameterized by metric space, pre-trained style-based visual search (SBVS). show first time that such space may be used to drive NST, enabling application interpolation styles from an SBVS system. The technical hyper-network predicts weight updates over diverse gamut content (portraits), tailoring parameterization per-region basis using semantic map facial regions. HyperNST exceed state art in preservation our stylized while retaining good performance.